نبذة مختصرة : M.Phil. ; Extensive studies have been conducted on cluster resource utilization due to the large investment on modern clusters. Although recent cluster schedulers and scheduling algorithms have significantly improved resource allocation by different architectural and algorithm designs, the allocated resources (i.e., containers) are often not adequately utilized by the demanding jobs due to over-estimated resource demands. In particular, we identify some dynamic resource utilization patterns that make the estimate of resource needs difficult even in common workloads. We show that existing scheduling techniques are ineffective for handling such dynamic resource utilization, as execution frameworks are treated as black boxes and resource usage is encapsulated in the container abstraction. ; To achieve high resource utilization for such workloads, we propose a new framework, called Ursa, which enables the scheduler to capture actual resource demands dynamically from execution runtime and to provide low-latency resource allocation. Ursa also allows executors to efficiently utilize the allocated resources. Experimental results show that Ursa significantly outperforms existing frameworks in terms of both makespan and average job completion time. ; 由於對現代集群的大量投資，針對集群資源的利用率有廣泛的研究。儘管最近的集群調度器和調度算法顯著改進了資源分配，但是由於過高估計的資源需求，所分配的資源（容器）通常不能被要求的作業充分利用。特別是，在常見的集群工作負載中往往出現這樣一種情況：任務對於資源的使用模式並不平穩，而是非常動態化的。我們表示現有的調度技術對於處理有大量這種任務的工作負載是缺乏效率的。 ; 為了實現此類場景中的高資源利用率，我們提出了一個名為Ursa的新框架，它使調度程序能夠從執行運行時動態捕獲實際資源需求，並提供低延遲資源分配。Ursa還允許計算執行框架有效地利用分配的資源。實驗結果表明，Ursa在整體工作量完成時間和平均工作完成時間方面均明顯優於現有框架。 ; Jin, Xiaoyue. ; Thesis M.Phil. Chinese University of Hong Kong 2019. ; Includes bibliographical references (leaves 49-55). ; Abstracts also in Chinese. ; Title from PDF title page (viewed on 25, November, 2020.
Processing Request
No Comments.